I build tools at the intersection of NLP × edge computing × news media.

My code solves one problem: ensuring factual accuracy when content crosses language boundaries and hardware platforms.

Deterministic QA for multilingual news — no AI, no hallucination, pure arithmetic.

• numlint — Catches magnitude errors, currency mismatches, and unit conversion bugs in translated text. 28 languages.
• ner-mundo — Cross-lingual named entity resolution. Merges partial names, translations, and coreferences into one canonical entry.
• Anseropolis — Linguistic fingerprinting × manipulation detection.
• TWGY — ASR correction for Taiwan Mandarin accents.

Linting tools for the systematic errors LLMs make when processing Chinese, Japanese, and Korean.

• kanjilint — CJK false friends (同形異義詞) detection. 116K entries, 1,025 trilingual-verified. First automated tool for this problem.
• world-media-tw-linter — Normalize 417 international media names in zh-TW text. Fixes LLM translation errors.
• president-linter — Detect stale world leader names (e.g. "南韓總統尹錫悅" → current). 35 G20+ leaders.

All three power inff.cc, a multilingual news aggregation pipeline.

Production NLP on RK3588S ($50 ARM board).

• GLiNER-Forge — Zero-shot NER across macOS (MLX), NVIDIA (CUDA), and RK3588S (ONNX INT8). Includes a fp16 quantization bug fix.
• rk-stantaz — Stanza NLP on RK3588S: 3.66× speedup. Found that RK3588S NPU is slower than CPU for LSTM workloads.
• rk-ckip — Traditional Chinese NLP (CKIP) on RK3588S: 3.32× total speedup.

• ckip-mlx — CKIP BERT on Apple MLX. WS/POS/NER for Traditional Chinese.
• ckip-coreml — Same, compiled for iOS/macOS via Core ML.
• mlx-gemma4 — MLX quantization for Google Gemma 4 with PLE-safe calibration.

• yabomish — A CJK input method for macOS, written in Swift.
• webplow — WebP conversion API in Go.
• glitchtip-mcp — MCP server for GlitchTip error tracking.
• oh-my-kiro — Behavioral tuning for Kiro CLI.
• inff.cc — A news aggregator I operate.

Mandarin · English · French